click through rate prediction

click through rate prediction

包括内容如下图：

 

使用直接估计法，置信区间置信率的估计：

1.使用二项分布直接估计

p(0.04<p^<0.06)=∑0.04n≤k≤0.06n(nk)0.05k0.95n−k p(0.04<p^<0.06)=∑0.04n≤k≤0.06n(nk)0.05k0.95n−k

1 2 3 4 5 6

low = ceil(n * 0.04 ); % 上取整 high = floor(n * 0.06 ); % 下取整 prob  =  0 ; for  i = low: 1 :high      prob  =  prob + nchoosek(n,i) * ( 0.05 ^i) * ( 0.95 ^(n - i)); end

2.使用正态分布近似

μ=p=0.05,σ2=p(1−p)n=0.05∗0.95n μ=p=0.05,σ2=p(1−p)n=0.05∗0.95n

1	normcdf( 0.06 , 0.05 ,sigma / x(i)^ 0.5 )  -  normcdf( 0.04 , 0.05 ,sigma / x(i)^ 0.5 )

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

warning off  all ; clear  all ;clc;close  all ; x = 500 : 1 : 1500 ; y  =  zeros( 1 ,size(x, 2 )); y2  =  zeros( 1 ,size(x, 2 )); sigma  =  sqrt( 0.05 * 0.95 ); for  i  = 1 :size(x, 2 )      y(i)  =  adPredict(x(i));      y2(i)  =  normcdf( 0.06 , 0.05 ,sigma / x(i)^ 0.5 )  -  normcdf( 0.04 , 0.05 ,sigma / x(i)^ 0.5 ); end   plot(x,y, 'b-' ); hold on; plot(x,y2, 'r-' ); hold on; x1 = [ 500  1500 ]; y1 = [ 0.85  0.85 ]; plot(x1,y1, 'y-' );

打印曲线：观测到，n=1000，差不多置信度会到达0.85

 

AUC概念及计算：

sklearn代码：sklearn中有现成方法，计算一组TPR,FPR，然后plot就可以；AUC也可以直接调用方法。

import numpy as np
import matplotlib.pyplot as plt

from sklearn.linear_model import LogisticRegression
from sklearn import datasets
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import roc_auc_score
from sklearn.metrics import roc_curve

digits = datasets.load_digits()

X, y = digits.data, digits.target
X = StandardScaler().fit_transform(X)

# classify small against large digits
y = (y > 4).astype(np.int)
X_train = X[:-400]
y_train = y[:-400]

X_test = X[-400:]
y_test = y[-400:]

lrg = LogisticRegression(penalty='l1')
lrg.fit(X_train, y_train)

y_test_prob=lrg.predict_proba(X_test)
P = np.where(y_test==1)[0].shape[0];
N  = np.where(y_test==0)[0].shape[0];

dt = 10001
TPR = np.zeros((dt,1))
FPR = np.zeros((dt,1))
for i in range(dt):
    y_test_p = y_test_prob[:,1]>=i*(1.0/(dt-1))
    TP = np.where((y_test==1)&(y_test_p==True))[0].shape[0];
    FN = P-TP;
    FP = np.where((y_test==0)&(y_test_p==True))[0].shape[0];
    TN = N - FP;
    TPR[i]=TP*1.0/P
    FPR[i]=FP*1.0/N



plt.plot(FPR,TPR,color='black')
plt.plot(np.array([[0],[1]]),np.array([[0],[1]]),color='red')
plt.show()

#use sklearn method
# fpr, tpr, thresholds = roc_curve(y_test,y_test_prob[:,1],pos_label=1)
# plt.plot(fpr,tpr,color='black')
# plt.plot(np.array([[0],[1]]),np.array([[0],[1]]),color='red')
# plt.show()

rank = y_test_prob[:,1].argsort()
rank = rank.argsort()+1
auc = (sum(rank[np.where(y_test==1)[0]])-(P*1.0*(P+1)/2))/(P*N);
print auc
print roc_auc_score(y_test, y_test_prob[:,1])